Model Recommendations —
Which model to point at which
/wb*command, and why. The runtime canon lives atfrontEnd/wbc-ui/core2/packages/wb-flow/templates/commands/model_recommendations; this file is the publishing edition — opinionated, with the why spelled out.
The user runs four model surfaces: the agent Pro (the AI agent, the AI agent, Haiku 4.5), Google One (the AI agent, the agent 3 Flash), Go subscription (the AI agent/Flash, Qwen 3.6 Plus, GLM-5.1, Gemma 4), and Zen (GPT 5 Nano + extras). Every recommendation below assumes that inventory.
The four roles, and which model wins each
| Role | What the work actually is | 🥇 Gold pick | Why this one |
|---|---|---|---|
| 🧠 Planner | Multi-step strategy, decomposition, holding many constraints in mind | the AI agent | Nothing else holds a 5-task plan together with this kind of structural rigour. Every /wbPlan should default to it. |
| ✅ Validator | Code-quality judgment, scoring, brutal-honest review | the AI agent (free) | Punches well above its tier on critique tasks — finds the unflattering issues other models politely skip over. The the agent pick (gold by raw quality) is reserved for security work, where stakes justify cost. |
| 🔨 Worker | Surgical code edits, refactors, debugger work | the agent the AI agent | The most precise instrument for "change this exact thing without disturbing anything else." the agent gets too creative; the agent occasionally over-edits. |
| 📋 Mechanical | Run a command, parse output, format a report | the agent 3 Flash | Templates and structured output are its happy path. Faster than Haiku for the same quality on this class of task. |
Per-command gold picks
The default unless you have a specific reason to override:
| # | Command | Gold pick | Why |
|---|---|---|---|
| 01 | /wbSetup | the agent 3 Flash | Mechanical bootstrap; no judgment required |
| 02 | /wbContext | the AI agent | Free, deep enough to actually re-read the package |
| 03 | /wbPlan | the AI agent | The one place to never skimp |
| 04 | /wbAudit | the AI agent | Brutally deep findings; cheaper than the agent for similar critique quality |
| 05 | /wbReview | the AI agent | "Big-thinker" style review — explains why a change is wrong, not just that it is |
| 06 | /wbTest | the agent V4 Flash | Fast mechanical report; no reasoning needed once tests pass/fail |
| 07 | /wbRelease | the agent the AI agent | workspace:* resolution + version bumping is exactly the kind of tedious-but-precise work it excels at |
| 08 | /wbPublish | the agent 3 Flash | Pure mechanical execution |
| 09 | /wbDeploy | the agent 3 Flash | Same — mechanical |
| 10 | /wbClean | the AI agent | Dead-code detection benefits from the deep-read |
| 11 | /wbLicense | the agent the AI agent | Surgical edits at gating points |
| 12 | /wbRefactor | the agent the AI agent | The gold standard. Don't substitute. |
| 13 | /wbDebug | the agent the AI agent | Reasoning + precision — the rare combo this requires |
| 14 | /wbDoc | Qwen 3.6 Plus | Docs and boilerplate are its happy path; saves the agent budget |
| 15 | /wbStandup | the agent 3 Flash | Reads reports, writes a digest — mechanical |
| 16 | /wbVision | the AI agent | The weakest command in the system; if you're going to use it, use the best brain |
| 17 | /wbBroadcast | the agent the AI agent | Voice and tone work — the agent writes more naturally than the agent for marketing copy |
| 18 | /wbGit | the agent 3 Flash | Conventional Commits is templated; mechanical |
| 19 | /wbSecure | the AI agent | The only role where paranoia justifies the cost — security findings under-call from cheaper models |
| 20 | /wbTranslate | GLM-5.1 | Best free multilingual; Qwen second |
| 21 | /wbToWBC | the agent the AI agent | Code transformation requires precision |
| 22 | /wbMonetize | the agent the AI agent | Tier-gating plumbing — same precision argument as /wbLicense |
| 23 | /wbActOn | the AI agent | Triage requires honest ranking; the agent is least likely to soften priorities |
| 24 | /wbCheck | the AI agent | Pre-flight quiz tests deep understanding; cheap models pretend to know |
| 25 | /wbTrack | the agent 3 Flash | Logging + summarization; mechanical |
| 26 | /wbNext | the agent the AI agent | Needs to read recent reports AND make a judgment call — the agent is the right balance |
the agent's edition (docs/) frames model choice as a tier-list ("Gold/Silver/Bronze, pick by budget"). That framing leaks one bad assumption: that the per-command pick is a budget question. It usually isn't.
For 22 of the 26 commands, the right pick is determined by role match, not subscription tier. /wbRefactor on the AI agent is worse than /wbRefactor on the AI agent not because the agent is cheaper, but because the agent over-edits on refactor tasks — it rewrites code that didn't need rewriting. The cost gap is incidental.
The four commands where budget legitimately enters the decision: /wbAudit, /wbDebug, /wbReview, /wbContext — all validators, all places where the AI agent is genuinely competitive with the agent/the agent and the free tier saves real money on long sessions.
The one mistake to avoid
Defaulting to the AI agent for everything because "the best model is the safe pick." the agent is the wrong tool for /wbSetup, /wbTest, /wbPublish, /wbDeploy, /wbStandup, /wbGit, /wbTrack, /wbDoc — eight mechanical commands where the cost gap buys you nothing. Worse, on those tasks the agent often over-thinks: it adds structure, asks clarifying questions, suggests improvements you didn't ask for. the agent Flash just does the thing.
The corollary: don't reach for free tiers on /wbPlan, /wbSecure, /wbDebug, /wbRefactor, /wbCheck. Those are the four places where the model quality directly determines whether the command's output is useful.
When to deviate from the gold pick
- Long context (>100K tokens): swap to the AI agent regardless of command. Its context recall genuinely outperforms the agent/the agent past the 100K mark.
- Burning Pro budget: swap mechanical-tier picks down to the agent V4 Flash or GPT 5 Nano. No quality loss on
/wbTest,/wbGit,/wbStandup. - Brand-new package, no
context.mdyet: the AI doesn't know the codebase. Use the agent for the first three commands (/wbSetup,/wbContext,/wbAudit) until the context files settle, then drop down.
Verdict
The "gold pick" table above is opinionated. Roughly 70% of agentic value comes from getting /wbPlan, /wbDebug, /wbRefactor, /wbSecure right — those four account for almost every non-trivial decision the system makes. Everything else can run on free or near-free tiers without measurable degradation. Pay for the four commands; save on the other 22.